The investigation of construction and clinical application of image recognition technology assisted bronchoscopy diagnostic model of lung cancer

Yihong Deng, Yuan Chen, Lihua Xie, Liansheng Wang, Zhan JuanFrontiers in Oncology 2022 · DOI

This study, recently published in Frontiers in Oncology, explores the intersection of deep learning and clinical pulmonology. Researchers developed and validated an image recognition technology (IRT) model designed to assist clinicians in diagnosing lung cancer during bronchoscopy.

TL;DR

Researchers trained a ResNet101-based deep learning model to identify and classify lung cancer lesions from bronchoscopy images. The model matched the diagnostic accuracy of senior clinicians (10+ years experience) while processing hundreds of images in seconds, demonstrating its potential as a high-efficiency "second opinion" in the operating room.

What Problem Does This Paper Address?

Lung cancer remains a leading cause of global mortality, and early, accurate diagnosis is critical for survival. Bronchoscopy—where a camera is inserted into the airways—is a primary diagnostic tool. However, interpreting these visual feeds is highly subjective.

While veteran pulmonologists have high accuracy, junior doctors and trainees often struggle to differentiate between malignant tumors and benign conditions like inflammation or tuberculosis. This "experience gap" can lead to missed diagnoses or unnecessary biopsies. The authors sought to create an objective, automated system to standardize diagnostic quality and reduce the workload on medical staff.

Methods

The team employed a supervised deep learning approach:

  • Architecture: They utilized ResNet101, a classic deep residual network. ResNet is particularly effective here because its "shortcut connections" allow for very deep architectures without the "vanishing gradient" problem that plagues other models during training.
  • Transfer Learning: The model was pre-trained on the ImageNet dataset to establish basic feature recognition before being fine-tuned on specialized bronchoscopic images.
  • Training Specifics: They used the Adam optimizer with a cross-entropy loss function. To handle data imbalance (specifically for adenocarcinoma), they applied data augmentation techniques like horizontal flipping.
  • Human-in-the-Loop Comparison: To validate clinical utility, the researchers invited nine clinicians—divided into senior, junior, and novice groups—to diagnose the same test set, providing a direct "man vs. machine" performance benchmark.

Datasets

The study relied on a retrospective dataset from the Third Xiangya Hospital (2015–2020):

  • Total Images: 2,238 lesion images.
  • Malignant Cases: 666 cases, including Squamous Cell Carcinoma (SCC), Adenocarcinoma (AC), and Small Cell Lung Cancer (SCLC).
  • Benign Cases: 152 cases, including infectious lesions, chronic inflammation, and tuberculosis.
  • Data Split: The data was partitioned into training, verification, and test sets at a 7:1:2 ratio.

Key Insights and Ideas

  • Bridging the Experience Gap: The most significant finding was that the model's performance was statistically equivalent to senior doctors. This suggests that AI can effectively "upload" a decade of clinical experience into a tool usable by a novice.
  • Efficiency at Scale: The model diagnosed 467 test images in just 30 seconds. In a clinical workflow, this speed allows for near-instantaneous feedback, which is vital during live procedures where patient sedation time must be minimized.
  • Morphological Nuance: Beyond simple "cancer vs. no cancer" detection, the model attempted to distinguish between pathological types (like SCC vs. AC). While less accurate than binary classification, it showed that the neural network is picking up on subtle textural and structural cues in the bronchial mucosa that are often invisible to the untrained human eye.

Key Results

  • Diagnostic Accuracy: For differentiating benign vs. malignant lesions, the model achieved an accuracy of 0.951 and an Area Under the Curve (AUC) of 0.940.
  • Sensitivity and Specificity: The model showed high sensitivity (0.978), meaning it rarely missed a malignant lesion, though its specificity (0.833) suggests a slightly higher rate of "false alarms" compared to its detection rate.
  • Pathological Classification: When tasked with distinguishing between SCC and AC, the model maintained a respectable overall accuracy of 0.745.
  • Consistency: The model’s agreement with senior doctors (Kappa = 0.658) was significantly higher than its agreement with novices (Kappa = 0.395), confirming it "thinks" more like an expert.

Limitations and Future Work

The authors identify several hurdles before this can be deployed in every hospital:

  1. Static vs. Dynamic: The model was trained on stationary images. However, bronchoscopy is a dynamic video procedure. Future iterations need to process real-time video feeds to account for motion blur and changing angles.
  2. Sample Imbalance: The number of benign cases (152) was significantly lower than malignant ones (666). Expanding the benign dataset is necessary to reduce false positives.
  3. Single-Center Bias: All data came from one hospital. To ensure the model isn't just "memorizing" the specific camera equipment or patient demographics of one facility, multi-center trials are required.
  4. Black Box Nature: While the model is accurate, the paper does not extensively use "explainable AI" (like Grad-CAM) to show which parts of the image triggered the diagnosis, a feature many clinicians desire for trust-building.

Summary updated

Abstract

The performance of our diagnostic model to distinguish benign and malignant lesions in bronchoscopy is roughly the same as that of experienced clinicians and the efficiency is much higher than manually. Our study verifies the possibility of applying IRT in diagnosis of lung cancer during white light bronchoscopy.
BibTeX
@inproceedings{deng2022investigation,
  author = {Yihong Deng and Yuan Chen and Lihua Xie and Liansheng Wang and Zhan Juan},
  title = {{The investigation of construction and clinical application of image recognition technology assisted bronchoscopy diagnostic model of lung cancer}},
  booktitle = {Frontiers in Oncology},
  year = {2022},
  doi = {10.3389/fonc.2022.1001840},
}
Figure